Does my understanding of GCN implementation right?

Questions
#1

From https://docs.dgl.ai/tutorials/basics/1_first.html

def gcn_message(edges):
    return {'msg' : edges.src['h']}

def gcn_reduce(nodes):
    return {'h' : torch.sum(nodes.mailbox['msg'], dim=1)}


class GCNLayer(nn.Module):
    def __init__(self, in_feats, out_feats):
        super(GCNLayer, self).__init__()
        self.linear = nn.Linear(in_feats, out_feats)

    def forward(self, g, inputs):

        g.ndata['h'] = inputs

        g.send(g.edges(), gcn_message)

        g.recv(g.nodes(), gcn_reduce)

        h = g.ndata.pop('h')

        return self.linear(h)

My question is about gcn_message and gcn_reduce.

If we pass all the edges into gcn_message, does it mean that each edge will send message from its own node?

If we pass all the nodes into gcn_reduce, does it mean that each node will recv the message from all of its edges?

if I edit gcn_reduce to

def gcn_reduce(nodes):
    recv_msg = nodes.mailbox['msg'] # [1,9,34]
    recv_msg2 = torch.sum(recv_msg, dim=1) # [1,34]
    return {'h' : recv_msg2}

I debug and stop at the first time. What does [1,9,34] mean in the code above?

Thank you.

#2

DGL did degree bucketing to accelerate the computation. The computation on nodes with same in-degrees are batched together. For example, if both node 1 and 2 have 3 in edges, the incoming message would be batched together, therefore the shape of nodes.mailbox['msg'] would be [2(bucket size, two nodes has same in degrees), 3(node’s degree is 3), feat_size].

#3

Thank you!

If we pass all the edges into gcn_message , does it mean that each edge will send message from its own node?

If we pass all the nodes into gcn_reduce , does it mean that each node will recv the message from all of its edges?

Am I right？

#4

I am sorry.

“both node 1 and 2 has 3 in edges”

Could you please upload a figure to describe this sentence?

Thank you!

#5

It just means node has 3 edges. Since DGLGraph is a directed graph, therefore I say 3 in edges (incoming edges).

#6

Thank you.

I understand.

Each node has 3 incoming edges and there exists 2 nodes.

I had better improve my English language.

#7

Hey all,

I have following this example as well, and have a question about it.

labeled_nodes = torch.tensor([0, 33]) # only the instructor and the president nodes are labeled
labels = torch.tensor([0, 1]) # their labels are different

Based on my understanding, those lines assume that only node 0 and 33 are known (I think because this example demonstrate semi-supervised technique). My question is what if all the nodes are known? Can I still modify this code to make it supervised technique?

Thanks in advance.

#8

The GCN example (and the original GCN paper) works under a semi-supervised setting, meaning that you have the entire graph beforehand but labels on only some of the nodes, and you need to predict the labels of the rest of nodes.

#9

Yes. You can use all labels if you have. However you may need to split train/validation/test set for generalization.

#10

Thanks @VoVAllen,

Let me give a try.

And to follow it up, once I have trained a model, save it, and load it again, how can I infer the nodes in a new graph? I tried to find a demo about it, but most of the demos usually perform train-test-eval in one go.

Thanks again